运行模型

要配置 sing-box 并运行模型,可以按照以下步骤进行：

步骤 1：安装 singbox

安装 singbox 库：

pip install singbox

步骤 2：创建配置文件

创建两个文件：singbox.py 和 config.json

singbox.py

from singbox importsingbox, TrainableModel, ImageLoader, ModelLoader
singbox(singbox_name="my-model", 
        data_set="coco", 
        labels="coco_labels.json",
        trainable_model_name="my-model-trained",
        trainable_model_load_from="coco_labels.json",
        trainable_model_lockup_iter=1,
        trainable_model_learning_rate=.1,
        trainable_model_batch_size=32,
        trainable_model_num_epochs=1,
        trainable_model_output_dir="output")
singbox.run_model()

config.json

{
    "singbox_name": "my-model",
    "data_set": "coco",
    "labels": "coco_labels.json",
    "trainable_model_name": "my-model-trained",
    "trainable_model_load_from": "coco_labels.json",
    "trainable_model_lockup_iter": 1,
    "trainable_model_learning_rate": 0.1,
    "trainable_model_batch_size": 32,
    "trainable_model_num_epochs": 1,
    "trainable_model_output_dir": "output"
}

步骤 3：下载训练好的模型

from singbox import download_coco
download_coco()

下载后,将模型文件加载到训练好的模型中：

from singbox import TrainingModel, ModelLoader
# 加载训练好的模型
TrainingModel.load_model(singbox_name="my-model-trained", model_load_from="output/my-model-trained.pth")
# 确保模型已经加载完成
TrainingModel.model()

步骤 4：定义函数进行图像数据读取

import cv2
import numpy as np
def read_image(image_path):
    image = cv2.imread(image_path)
    if image is None:
        raise ValueError(f"Could not read image '{image_path}'")
    # 归一化处理
    image = (image / 255.) - 0.5
    return image, image_path
# 示例：读取图像并进行预测
image_path = "path_to_image.jpg"
image, image_path = read_image(image_path)
_, predicted = singbox.model.predict(image)
# 输出预测结果
print(f"预测结果：{predicted}")

步骤 5：定义模型预测函数

def predict_and_calculate_accurate_rate(image_path, labels):
    image, image_path = read_image(image_path)
    predicted = singbox.model.predict(image)
    # 分类标签
    accurate_count = 0
    total_count = len(predicted)
    for i in range(total_count):
        label = labels[i]
        if label == predicted[i]:
            accurate_count += 1
    if total_count == 0:
        return 0.
    else:
        return accurate_count / total_count
# 示例：计算准确率
labels = ["cat", "dog", "bird"]
accuracy = predict_and_calculate_accurate_rate("cat.jpg", labels)
print(f"准确率：{accuracy}")

步骤 6：运行模型

accuracy = predict_and_calculate_accurate_rate("test.jpg", labels)
print(f"准确率：{accuracy}")

步骤 7：输出结果

print("模型运行结果：", accuracy)

注意事项

1. 配置文件的正确性：确保配置文件中的所有参数正确，包括模型名称、数据集、标签文件等。
2. 路径的正确性：确保图像路径和标签文件的文件名和路径与配置文件中的数据一致。
3. 模型加载：确保加载模型时参数正确,特别是模型文件的路径和文件名。
4. 模型训练：如果模型未加载,可能需要重新训练模型。
5. 错误处理：如果模型运行失败,检查是否有错误信息并尝试重新加载模型或调整参数。

通过以上步骤，您可以成功配置singbox并运行模型,获得准确率和结果。